Fix datafusion-cli memory leak by using snmalloc

[2010YOUY01]

Which issue does this PR close?

• Closes Memory held on in datafusion-cli (reported memory size increases over time) #15939

Rationale for this change

To be more precise: this is not a memory leak, instead, it's memory retention by the global allocator.

The issue is in datafusion-cli's REPL, the default mimalloc global allocator will hold memory between queries, and let the memory usage accumulate. I think it will only free the allocation lazily when hitting the physical memory limit.
It won't cause OOM errors, but I've observed that some memory-intensive queries slow down when run after the global allocator has retained a significant amount of memory.

The direct fix is to call the purge function https://microsoft.github.io/mimalloc/group__extended.html#ga421430e2226d7d468529cec457396756, however, it's tricky to be done (has to be applied once a while on all worker threads).

I tried snmalloc, and the result shows:

• No more memory retention issues
• Some memory-consuming queries run faster and have a smaller memory footprint:

-- This query runs 50% faster and uses 30% less memory with `snmalloc`
SELECT *
FROM generate_series(1, 10000) AS t1(v1)
JOIN generate_series(1, 100000) AS t2(v2)
ON (t1.v1 + t2.v2) % 10 = 0;

snmalloc is also one of the fastest allocators like mimalloc. Additionally, the DataFusion benchmark already supports snmalloc, so I believe switching to snmalloc is a viable fix.

What changes are included in this PR?

Uses snmalloc by default in datafusion-cli, mimalloc is still kept as a feature.

Are these changes tested?

Tested manually for the reproducer in #15939

Are there any user-facing changes?

[Omega359]

Any reason for snmalloc vs jemalloc?

[2010YOUY01]

Any reason for snmalloc vs jemalloc?

Experiments show snmalloc is faster: #1264

[Dandandan]

I think it makes sense to run some benchmarks again since also jemalloc has implemented some optimizations over time.

[Dandandan]

I see the benchmarks currently are actually using mimalloc, so can you change them as well, so we can also see whether those might be regressing @2010YOUY01 ?

[Dandandan]

Snmalloc is a bit slower on average for me, although keeping too memory reserved is problematic as well:

┏━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Query        ┃       main ┃ test_snmalloc ┃        Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ QQuery 0     │     1.20ms │        1.22ms │     no change │
│ QQuery 1     │    23.89ms │       23.75ms │     no change │
│ QQuery 2     │    68.23ms │       68.67ms │     no change │
│ QQuery 3     │    53.47ms │       73.36ms │  1.37x slower │
│ QQuery 4     │   508.55ms │      537.01ms │  1.06x slower │
│ QQuery 5     │   632.27ms │      782.97ms │  1.24x slower │
│ QQuery 6     │     1.13ms │        1.15ms │     no change │
│ QQuery 7     │    28.45ms │       26.77ms │ +1.06x faster │
│ QQuery 8     │   557.34ms │      657.69ms │  1.18x slower │
│ QQuery 9     │   796.61ms │      899.81ms │  1.13x slower │
│ QQuery 10    │   189.65ms │      200.76ms │  1.06x slower │
│ QQuery 11    │   210.78ms │      214.56ms │     no change │
│ QQuery 12    │   705.93ms │      822.20ms │  1.16x slower │
│ QQuery 13    │   960.91ms │     1064.18ms │  1.11x slower │
│ QQuery 14    │   592.73ms │      636.50ms │  1.07x slower │
│ QQuery 15    │   645.67ms │      569.91ms │ +1.13x faster │
│ QQuery 16    │  1465.49ms │     1389.75ms │ +1.05x faster │
│ QQuery 17    │  1236.71ms │     1298.29ms │     no change │
│ QQuery 18    │  2313.13ms │     2813.46ms │  1.22x slower │
│ QQuery 19    │    49.86ms │       46.93ms │ +1.06x faster │
│ QQuery 20    │   941.72ms │     1012.69ms │  1.08x slower │
│ QQuery 21    │  1110.70ms │     1171.63ms │  1.05x slower │
│ QQuery 22    │  1707.81ms │     1897.32ms │  1.11x slower │
│ QQuery 23    │  6344.35ms │     7340.64ms │  1.16x slower │
│ QQuery 24    │   361.13ms │      403.85ms │  1.12x slower │
│ QQuery 25    │   364.46ms │      382.83ms │  1.05x slower │
│ QQuery 26    │   432.50ms │      462.67ms │  1.07x slower │
│ QQuery 27    │  1341.14ms │     1421.37ms │  1.06x slower │
│ QQuery 28    │ 11386.96ms │    11408.64ms │     no change │
│ QQuery 29    │   440.27ms │      496.50ms │  1.13x slower │
│ QQuery 30    │   635.64ms │      617.48ms │     no change │
│ QQuery 31    │   622.02ms │      668.88ms │  1.08x slower │
│ QQuery 32    │  2610.37ms │     2290.81ms │ +1.14x faster │
│ QQuery 33    │  2726.09ms │     3222.04ms │  1.18x slower │
│ QQuery 34    │  3187.37ms │     3250.72ms │     no change │
│ QQuery 35    │   943.58ms │      889.43ms │ +1.06x faster │
│ QQuery 36    │    82.98ms │       80.34ms │     no change │
│ QQuery 37    │    37.27ms │       35.97ms │     no change │
│ QQuery 38    │    83.25ms │       81.83ms │     no change │
│ QQuery 39    │   135.20ms │      130.03ms │     no change │
│ QQuery 40    │    28.83ms │       26.85ms │ +1.07x faster │
│ QQuery 41    │    27.56ms │       26.54ms │     no change │
│ QQuery 42    │    23.81ms │       22.27ms │ +1.07x faster │
└──────────────┴────────────┴───────────────┴───────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Benchmark Summary            ┃            ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ Total Time (main)            │ 46616.99ms │
│ Total Time (test_snmalloc)   │ 49470.29ms │
│ Average Time (main)          │  1084.12ms │
│ Average Time (test_snmalloc) │  1150.47ms │
│ Queries Faster               │          8 │
│ Queries Slower               │         21 │
│ Queries with No Change       │         14 │
└──────────────────────────────┴────────────┘

[github-actions]

Thank you for your contribution. Unfortunately, this pull request is stale because it has been open 60 days with no activity. Please remove the stale label or comment or this will be closed in 7 days.